NVIDIA: Llama 3.3 Nemotron Super 49B V1.5

Supports function calling via structured JSON schemas with native integration for tool definitions, enabling agents to invoke external APIs and functions with type-safe argument binding. The model was post-trained specifically for agentic workflows, allowing it to parse tool schemas, select appropriate functions, and generate properly-formatted invocation payloads without hallucination of non-existent tools.

Processes and reasons over retrieved documents injected into the context window, using the 128K token context to maintain long document chains and conversation history simultaneously. The model was post-trained on RAG-specific tasks, enabling it to synthesize information across multiple retrieved passages, cite sources implicitly, and distinguish between retrieved context and training knowledge.

Generates multi-step mathematical proofs and derivations with explicit reasoning chains, trained on mathematical problem-solving datasets to produce intermediate steps, symbolic manipulation, and formal reasoning. The model can handle algebra, calculus, linear algebra, and discrete math problems by decomposing them into verifiable steps rather than jumping to answers.

Generates and completes code across multiple programming languages (Python, JavaScript, Java, C++, etc.) with context-aware suggestions based on surrounding code, imports, and function signatures. Post-trained on code-specific tasks, the model understands language idioms, common libraries, and can generate both snippets and full functions with reasonable correctness.

Synthesizes scientific knowledge across physics, chemistry, biology, and related domains, generating explanations grounded in scientific principles and literature. Post-trained on science-specific reasoning tasks, the model can explain mechanisms, predict outcomes, and reason about experimental design with domain-appropriate terminology and accuracy.

Maintains coherent multi-turn conversations with up to 128K tokens of context, enabling long document discussions, extended reasoning chains, and conversation history preservation without context truncation. The model can reference earlier turns, maintain character consistency, and reason over accumulated context without losing track of prior statements.

Follows complex, multi-step instructions by decomposing tasks into subtasks, maintaining task state across turns, and executing instructions with high fidelity to user intent. The model can handle conditional logic, iterate on feedback, and adapt execution based on intermediate results without losing track of the original goal.

Primarily optimized for English with capability to understand and translate from other languages into English, leveraging Llama-3.3's multilingual foundation while maintaining English-centric post-training. The model can process non-English input and translate to English for reasoning, then generate English responses, though non-English output quality is not guaranteed.

Achieves 49B parameter efficiency through knowledge distillation from the larger Llama-3.3-70B-Instruct model, maintaining reasoning capability and instruction-following quality while reducing inference latency and memory requirements. The distillation process preserves agentic workflow performance through careful SFT on tool-calling, RAG, and reasoning tasks.